Speakers
Description
Despite the growing availability of digital dictionaries, semantic interoperability often remains limited to headwords, metadata, or article structures. Dictionary senses remain difficult to compare because they are shaped by different editorial traditions, segmentation practices, and levels of semantic granularity. This paper asks whether taxonomy-guided LLM classification can generate reliable, reviewable candidate links between heterogeneous dictionary senses. We test this approach on sense-level units from the letter range N in four German dialect dictionaries: Mecklenburgisches Wörterbuch, Hessen-Nassauisches Wörterbuch, Pfälzisches Wörterbuch, and Schweizerisches Idiotikon. Using GPT-5 via the KISSKI interface, each unit is assigned ten times to Post’s onomasiological taxonomy, which serves as a shared conceptual reference model. Evaluation combines exact agreement with reference labels, hierarchical proximity, run-to-run reproducibility, and majority aggregation. Exact correctness ranges from 74.91% to 80.88%; top-level taxonomic agreement exceeds 91%; and Fleiss’ κ values between 0.78 and 0.83 indicate substantial reproducibility. Majority aggregation raises exact correctness to 78.61–84.28%. The results show that taxonomy-guided LLM classification can support sense-level interoperability by producing interpretable candidate links and reliability signals for targeted lexicographic review.